Busy people would benefit from ambient, adaptive tools for reminding them what they have to do, depending on various contextual parameters. We are developing an adaptive and expressive agent, which learns when and how to notify users about self-assigned tasks and events. In this paper, we focus on two crucial issues for such a system: the selection of the input data set, and the design of an appropriate mechanism to get user feedback without being too intrusive. We describe in particular the inputs that encapsulate the current context of the agent: relative temporal distances, historical information about reminders and categories, and the context of both the user and the hosting device.